About - ANR Robot

On this site I use 'we', partly because I am hoping to bring in collaborators, and also because I thought it sounded better on a website. For now, ANR Robot is just me, an independent research scientist.

My interest in robotics probably begins in childhood with an interest in electronics — crystal radio, Heathkit TV — and in sports cars and mechanics, but also in how the brain works, in psychology, and in computers, so I guess I was interested in AI. My undergraduate degree was in computer science and electrical engineering, but then I veered off a little into theoretical physics for my PhD. Eventually, I came back to AI through an overlap between physics and neural networks, which led me to co-found a face-recognition company — Visionics. Since then I have come full circle to robotics, which pretty much combines every interest.

Current Research Focus

Today, robots can dance and AIs can think and plan and understand what they see, but robots can't yet autonomously hand-embroider or assemble a ship model from a kit! We can't give a robot mechanical and electrical components, tools, a bench, and previously unseen written instructions and expect the robot to build a robot. In our opinion, combining human-level dexterity with human-level planning and understanding is at the frontier of robotics. This is where we are focusing our attention.

As a small lab, we must use all resources available, and so we are testing, tweaking, and modifying open-source software and off-the-shelf hardware. The idea is to be as close to the state-of-the-art as possible on a reasonable budget and with a very small team (one for now!). What we can't do, at the moment, is pre-train a robot on a huge dataset, so we try to leverage models which come with pre-trained weights, such as pi0.5 and GR00T and other VLAs, and also models such as ACT and its descendants, which do not require pre-training. In addition to VLA 'policy' models, which live in the moment, we also plan to try out world models such as NVIDIA's Cosmos, since it is likely that robots need to reason about the future too.

For hardware, we are using the very robust and reasonably priced Trossen Stationary AI robot, to which we are planning to add various end-effectors — hands — to improve dexterity.

Collaboration & Contact

I am actively interested in talking with people working on similar problems — in academia, corporate labs, robotics companies, and hobbyists. I am happy to share datasets, code, or just trade notes.